Abstract
Making computer-generated (CG) images more difficult to detect is an interesting problem in computer graphics and security. While most approaches focus on the image rendering phase, this paper presents a method based on increasing the naturalness of CG facial images from the perspective of spoofing detectors. The proposed method is implemented using a convolutional neural network (CNN) comprising two autoencoders and a transformer and is trained using a black-box discriminator without gradient information. Over 50% of the transformed CG images were not detected by three state-of-the-art spoofing detectors. This capability raises an alarm regarding the reliability of facial authentication systems, which are becoming widely used in daily life.
Introduction
Facial authentication systems are ubiquitous, securing everything from personal mobile devices to financial transactions. To defend against presentation attacks—where an attacker uses a photo, video, or synthetic image—these systems employ spoofing detectors (liveness detection). While human observers are increasingly adept at identifying "uncanny valley" artifacts in computer-generated (CG) faces, automated detectors rely on subtle statistical inconsistencies in texture and edge distributions that distinguish synthetic pixels from organic skin.
Traditional computer graphics research focuses on improving the visual realism of CG images for human perception during the rendering phase. This study explores a different approach: a post-rendering transformation designed specifically to deceive machine-learning-based spoofing detectors. By bridging the gap between synthetic and natural image distributions, the authors demonstrate that even sophisticated detectors can be bypassed.
Methods
The core of the proposed approach is H-Net, a convolutional neural network (CNN) characterized by its H-shaped architecture. This system is designed to map features from the CG domain into a "natural" latent space.
H-Net Architecture
The H-Net consists of three primary components:
- Dual Autoencoders: The network utilizes two separate autoencoders—one trained on natural images and the other on CG images. This allows the system to learn the fundamental latent features of both domains.
- Transformer Network: Positioned between the two autoencoders, the transformer's role is to map the latent representations of CG images onto the latent feature space of natural images.
- Black-Box Training Strategy: Since many commercial spoofing detectors are "black boxes" that do not provide gradient information, the authors utilized a VGG-based perceptual loss to approximate the necessary gradients. This enables H-Net to learn "naturalization" without internal access to the target detector.
The training process involves four sequential steps to refine the autoencoders and the transformer. Once trained, a CG input is processed through the CG encoder, translated by the transformer, and reconstructed by the natural decoder to produce a "naturalized" output.
Datasets
The performance of H-Net was evaluated using three datasets representing varying levels of CG quality:
- Dataset 1: 240 high-realism CG images (from web sources and the PES 2012 game) and 240 natural images.
- Dataset 2: 270 3D face scans from the Basel Face Model and 270 images from the Caltech Faces 1999 dataset.
- Dataset 3: 3,236 grayscale CG images from the MIT CBCL dataset and 3,236 natural images from MS-Celeb-1M.
Results and Discussion
The H-Net transformation proved highly effective at deceiving biometric security measures. Across three state-of-the-art spoofing detectors, over 50% of the transformed CG images successfully avoided detection.
Key findings include:
- Detection Rate Reduction: Against the "Wu et al." detector, the detection rate for Dataset 1 dropped from 93.33% to 0.00% following the H-Net transformation.
- Identity and Feature Preservation: By performing transformations in the latent space, H-Net maintains the subject's identity, facial expression, and lip shape. This ensures the image passes the subsequent face recognition phase of the authentication pipeline.
- Visual Naturalization: Qualitatively, H-Net improves skin textures and normalizes color and brightness, effectively removing the "flatness" typical of CG renders.
Conclusion and Future Work
The success of H-Net highlights a significant vulnerability in current facial authentication systems. However, several challenges remain. The authors noted a "black skin-color" problem where training data bias led to undesirable whitening of darker-toned CG faces. Additionally, the current model is limited to 256x256 static images. Future research will focus on higher resolutions and real-time video transformations, which are necessary to bypass liveness checks requiring dynamic movements like blinking or speaking.
Summary updated